My report from the TC39 conference
Last week, I attended an event that may seem somewhat unlikely for an interpreter: the 39th “Translating and the Computer” conference in London. Yes, you read that right, AsLing - the International Association for Advancement in Language Technology - has been organising this conference for almost four decades now. Anyone who has ever had even a small role in organising an annual event will realise what a truly remarkable achievement this is.
A few months ago, my friend and colleague Anja Rütten told me about this conference. She had attended before and knew the organisers were seeking to branch into interpreting more. So we got a handful of colleagues together and worked on the idea of organising a panel discussion at the conference. Given the organiser and the topic of the conference, the topic was a no-brainer: What does the future in general, and language technology in particular, hold for interpreters?
Fittingly, the entire preparation took place online: the abstract was drafted jointly in a Google Doc, the outline for the panel came together in an online spreadsheet - just what the doctor (Rütten) ordered.
Last week, all the panelists descended upon London:
- Dr. Anja Rütten
- Joshua Goldsmith (tablet interpreting partner-in-crime)
- Barry Slaughter Olsen, the tech-savvy interpreter,
- Marcin Feder, also a tablet interpreter and currently in charge of technology projects at the European Parliament’s interpreting service,
- our moderator Danielle D’Hayer, associate professor for interpreting at London Metropolitan University.
The venue for this conference was spectacular: One Birdcage Walk is located in the home of the Institution of Mechanical Engineers, very close to the Houses of Parliament and right next to St. James’ Park. During the conference, we used the big lecture hall (under the watchful eye of the “father of railways”, George Stephenson.
Due to my travel arrangements, I missed the opening remarks by AsLing president João Esteves-Ferreira. But I did arrive in time for Roberto Navigli’s opening keynote about the ever-growing and fascinating universe of BabelNet. In the simplest terms, BabelNet is a multilingual online reference tool that integrates many resources such as WordNet, Wikipedia, Wiktionary and GeoNames to provide both a dictionary and a semantic network of terms. BabelNet currently covers almost 300 languages and closes gaps by using machine translation. The caveat: many resources are created automatically, and thus need to be verified by a qualified human. Related tools that make use of BabelNet are Babelfy (a tool that provides context for entered text, across several languages) and eRead (an upcoming software to enhance the electronic reading experience).
After the first coffee break (or “health break”, as AsLing calls them), Andrzej Zydroń looked “Beyond Neural Machine Translation”. This latest step change in machine translation has been getting lots of praise recently for its improved quality and is used by technology giants like Google and Facebook. Zydroń pointed out, however, that while NMT scores better for fluency or morphology, it does not necessarily provide better accuracy. In fact, improved fluency can make it harder to spot mistakes! NMT is also more sensitive to typos and grammatical errors in the source text and does worse with sentences longer then 60 words. And since NMT is basically a black box, tuning the algorithms is nigh impossible. So unless we want to limit NMT for “gisting”, Zydroń suggested, we should use “self-learning” neural machine translation that integrates dynamic feedback (as used by Lilt or SDL Adaptive MT).
Carlos Teixeira then presented a very interesting experiment in machine interpreting: His team at Dublin City University created a simple machine translation tool (based on BabelNet) to research how different input methods (keyboard/mouse/touch/speech recognition) influence the work of translators. To pick out just one aspect: translators were more productive with keyboard-and-mouse input and also preferred this input method to the other ones tested. Dictation came second in user preference. Teixeira and his team intend to continue working on the browser-based MT tool, integrating more features and making it available commercially.
In the last talk of the morning session, Kristiina Abdallah from Jyväskulä University shared the less-than-ideal experience of Finnish audiovisual translators who made the subtitles for the “Star Wars: The Force Awakens” film. Due to the many limitations that come with a blockbuster production - secrecy being just one of them - the translators were not happy with their work and explicitly declined to be named in the film credits.
The afternoon part started with two short poster presentations:
- Laura Cacheiro Quintas showed us the “Miro Translate” subtitling tool for educational material. Developed at the Perpignan Via Domitia University, Miro Translate uses the Microsoft Translator Speech API for automatic speech recognition (ASR) and automatic captioning with subsequent post-editing.
- Marianne Starlander (University of Geneva) described how she adapted her MA course about CAT tools to new trends. More and more enrolled students require new teaching methods such as crowdsourcing or online quizzes. And the integration of MT into CAT tools makes it necessary to rethink how students can be familiarised with those tools.
Natalia Bondonno from UNDGACM in New York introduced the eLUNa family of language tools used by translators at the United Nations. eLUNa includes a web-based CAT tool, an editorial interface and a search engine, all specially designed for (and using input from) UN language professionals.
My highlight of the day was definitely the talk by Sarah Griffin-Mason, the current president of the Institute of Translators and Interpreters in the UK. It had a very keynote-like vibe and discussed the future of translators until 2045 and beyond. Sarah referenced the work of many experts in technology and artificial intelligence, like Ray Kurzweil, Olle Haggström or Nicholas Carr (who recently published “The Glass Cage”).
Much of what Sarah said resonated with me and the rest of the audience:
- “We’re at the stage where we see the wheel but don’t know what life with the wheel is like. Not to speak off self-driving cars.”
- “Let’s reclaim the joy in our technology!”
- “It’s not about human versus machine, but human plus machine.”
- “Technology increases productivity and can give us better tools, but human judgement will and must be key.”
The last point on the agenda for Thursday was a round table about the future of the translator (not translation), which, I’m sorry to say, I found less interesting than Sarah’s talk before.
The small band of interpreters got together after the conference in a nearby bar to discuss the panel on the next day. Needless to say, we had a great time. (Which then continued at the conference networking dinner at the National Theatre.)
Friday was off to an excellent start with Anja Rütten and her overview of terminology tools for interpreters. All the panelists were there and it was a really good session. Interestingly, almost all the tools presented by Anja are developed by interpreters or relatives/friends/spouses of interpreters. And just like interpreters, they are all somewhat idiosyncratic in how they work. Some tools are no longer actively maintained, which underlines how much of a niche market interpreting is.
The Friday keynote came from someone whose work I’ve been following for years: Professor Alexander Waibel. He has been working on “A World Without Language Barriers” (the keynote title) for decades and relayed that history in the first part of this talk. Waibel also pointed out that the barrier for communication today is no longer technology - illustrated by the astronomic rise of mobile phone users worldwide - but rather: language. He took a deep dive into the inner workings of neural networks and how they can be leveraged for spoken language translation and even for recognising emotion in speech. (Note that Waibel doesn’t speak of machine interpreting!) The “long tail of languages” was also an intriguing concept to me - how can we scale language technology from a handful of “big” languages to the thousands of other languages in the world? If you’d like to know more about his many projects, including the “Lecture Translator” he has been using since 2005 for automatic captioning and translation of his lectures at the Karlsruhe Institute of Technology, check out his KIT page.
Continuing the interpreting track, Claudio Fantinuoli introduced the new ASR (automatic speech recognition) feature in his interpreter terminology application InterpretBank. Watch the video embedded here, it’s truly impressive. He also established the new (to me) category of “computer-assisted interpreting” (CAI) - very useful! Fantinuoli stressed the idea of “cognitive saturation” in interpreting. CAI tools can help us improve (and speed up) preparation and access terminology quickly. Automatic speech recognition is one way to further reduce the cognitive load of technology and can provide the interpreter with terms from the glossary (plus translations), numbers and named entities (such as names or organisations). InterpretBank’s ASR engine uses Dragon Naturally Speaking (offline on the computer) or the Microsoft Bing speech recognition API (online).
Bianca Prandi continued the “Italians working at Germersheim University” two-punch. She researches how interpreters use CAI tools since process-oriented interpreting studies publications about cognitive load are far and few between. One model that can be used is Kilian Seeber’s “Model of cognitive load in simultaneous interpreting”. However, given the current lack of data, it’s difficult to have a proper conversation about this issue. With her PhD, Bianca is working to change that. And she uses sophisticated methods like eye-tracking, pupillometry and key-logging for target-text and terminology analysis and to detect variations in cognitive load. The results are mapped onto a conflict matrix.
If you’re curious, there’s a recorded webstream of a joint presentation by Bianca and Claudio available on the FTSK website.
Josh Goldsmith and I have been collaborating on tablet interpreting for a while, so it was great to see him present the findings of his latest research on how interpreters use tablets for consecutive interpreting. He compared and contrasted the different tablets and accessories currently available on the market. His conclusions provided an interesting insight into how “tablet interpreters” use technology for their work and can also serve as a useful guide to allow colleagues to pick the devices, applications and styluses which best meet their needs.
Next up was our panel! Here's the full recording:
Day two wrapped up with a talk on how machine translation can be used for Bible translation. The problem when using MT for Bible translation, apparently, is that both statistical and neural MT systems need large amounts of training data to work properly. For smaller (or “low-resource”) languages, however, that training data usually does not exist. This talk demonstrated an approach to solve that problem. It was above my pay grade, but still entertaining to listen to.
And then, very soon, very suddenly, the conference was over! I had a fantastic time in London (despite the bad wifi…) and would definitely consider attending again.